Geosatial evaluation has develop into a necessary instrument for monitoring land use, agriculture, and environmental modifications. One of the highly effective platforms for large-scale geospatial processing is Google Earth Engine (GEE). Final yr, I labored on a mission that concerned detecting sugarcane plantations utilizing distant sensing knowledge and inbuilt machine studying fashions in GEE.
On this article, I’ll stroll you thru:
1. How I used GEE to detect sugarcane fields
2. The datasets and classification methods used
3. Some challenges and key takeaways from the mission
Let’s dive in!
GEE is a cloud-based geospatial platform that permits customers to investigate satellite tv for pc imagery and vector datasets at scale. Not like conventional GIS software program, it eliminates the necessity for heavy native processing.
Pattern of what GEE can do:
I additionally share GEE tutorials on YouTube, which you’ll try right here: https://youtube.com/playlist?list=PLKmAxy6Caj2DZNeXEdP9cKMYV1RhK0Wbx&si=BEhPo1v87NT77Ccq
I plan to revamp these tutorials with clearer explanations and higher audio, keep tuned! Lets proceed…
Key advantages of GEE:
– Entry to huge satellite tv for pc imagery datasets (e.g., Landsat, Sentinel-2)
– Cloud computing energy for large-scale processing
– Constructed-in machine studying algorithms for classification duties
For this mission, I primarily used:
- Pattern Drone Imagery
- Floor reality knowledge (discipline knowledge that had been collected then I added different courses for use for coaching)
So we first began by importing the Imagery To GEE by way of the belongings part from the place you’ll be able to add Raster Imagery (Geotiff, tiffs) or Vector knowledge (shapefiles and so forth):
For those who’re occupied with testing this, try the GEE script beneath if you want to go straight forward to the GEE Code hyperlink:
https://code.earthengine.google.com/9ef6d24cc33c5a1703154588b9f0de66
We then load that imagery asset now in our code editor:
Subsequent factor was getting NDVI which i’ll now then use in my evaluation after which selecting the polygons that i’ll use for classification because the coaching knowledge.
Tip: For those who’re utilizing GEE for vegetation evaluation, the NDVI (Normalized Distinction Vegetation Index) is a robust metric.
For classification, I used Random Forest (RF) and Assist Vector Machines (SVM), highly effective supervised studying algorithms, to categorise sugarcane based mostly on spectral traits.
- Outline coaching areas and break up into coaching and testing units (labeled sugarcane and different courses within the picture):
Code :
var training_points = sugarcane.merge(forest).merge(builtup).merge(bareland);
print(training_points, 'training_points');var Bands_selection = ["b1", "b2", "b3"];
var coaching = tiffImage.choose(Bands_selection).sampleRegions({
assortment: training_points,
properties: ['landcover'],
scale: 30
});
print(coaching, "coaching");
var bounded = training_points.geometry().bounds();
var filtered = training_points.filter(ee.Filter.bounds(bounded));
Export.desk.toAsset({
assortment: filtered,
description: 'SugarCaneAITrainingPointsAsset',
assetId: 'SugarCaneTrainingPointsAsset'
});
Export.desk.toDrive({
assortment: filtered,
description: 'SugarCaneAITrainingPointsCSV',
fileFormat: 'CSV'
});
Export.desk.toDrive({
assortment: filtered,
description: 'SugarCaneAITrainingPointsShapefile',
fileFormat: 'SHP'
});
// SPLITS: Coaching (75%) & Testing samples (25%)
var Total_samples = coaching.randomColumn('random');
var training_samples = Total_samples.filter(ee.Filter.lessThan('random', 0.75));
print(training_samples, "Coaching Samples");
var validation_samples = Total_samples.filter(ee.Filter.greaterThanOrEquals('random', 0.75));
print(validation_samples, "Validation Samples");
- Use the bands
- Prepare the Random Forest mannequin
- Classify the complete area
GEE Code Snippet for Sugarcane Classification half:
Visualizing the Outcomes
To visualise categorised areas:
1. Add to map by way of the tactic — Map.addLayer()
2. Export the outcomes as GeoTIFF or SHP for additional GIS evaluation
I additionally added key map parts such because the legend, title and did space estimation.
Like every distant sensing mission, there have been challenges:
🔸 Spectral confusion — Distinguishing between Crops eg. maize and sugarcane was troublesome resulting from their comparable spectral signatures. getting to truly know that is precisely sugarcane…
🔸 Floor reality knowledge limitations — Extra discipline knowledge typically results in higher mannequin accuracy.
For this mission, we used aerial imagery . Whereas the decision was considerably higher than satellite tv for pc imagery, there have been nonetheless challenges in precisely differentiating maize from sugarcane resulting from their spectral similarities. The picture supplied some enchancment, nevertheless it nonetheless wasn’t excellent.
Subsequent Steps: A extra sturdy resolution might contain incorporating deep studying to extract distinctive spectral and texture-based / sample options, enabling higher crop differentiation. That is one thing i’ll discover intimately and make use of the deep studying methods to provide you with exact outcomes for detection.
After refining the mannequin, we achieved 85% accuracy in sugarcane detection. The code is obtainable in my GitHub repository:
GitHub Repo: https://github.com/KimutaiLawrence/GEEProcessess
Direct GE Code hyperlink: https://code.earthengine.google.com/9ef6d24cc33c5a1703154588b9f0de66
Thanks!
My profiles:
Geospatial Information Scientist / Full-Stack GIS Developer Profile: https://github.com/KimutaiLawrence/KimutaiLawrence.github.io
YOUTUBE Channel: https://www.youtube.com/@geospatialprimetech997
Twitter: https://twitter.com/lawrence_kim_
Whatsapp Contact: +254759629059
Electronic mail: geospatialprime@gmail.com
